LLM cost optimization in Philadelphia.

The technical reality of institutional spend in Philadelphia.

Questions from Philadelphia buyers.

Our payer-provider correspondence pipeline is our biggest AI line. Where does the spend hide?

In three places the compliance-first build never revisited. Context re-billing: medical-policy excerpts, approved appeal language, and template libraries riding uncached on every case, paying full price thousands of times a day for content that never changes; restructured as PHI-free cached prefixes, that content bills at the 90-percent discount. Model oversizing: template-grounded drafting and correspondence classification running on frontier models when mid-size models match, provable on an eval set your appeals specialists grade. Synchronous defaults: overnight correspondence queues and next-day appeal drafts paying real-time rates when batch endpoints inside the same BAA umbrella cost half. None of the fixes touch the posture: audit logging, BAA channels, and specialist review stay exactly as your compliance office approved them, and the report documents that alongside the per-case savings, which typically run 60 to 80 percent against unoptimized baselines.

Can document review for litigation run at batch economics, and what does that change?

Yes, and it changes whether large-matter AI review is affordable at all. First-pass review work (issue tagging, chronology extraction, near-duplicate clustering, privilege-screen support) is latency-insensitive by nature: the review team works queues, not live responses, so the entire pipeline belongs on batch pricing at roughly half synchronous rates. The structure for a large matter: a cheap first tier (small model, cached protocol instructions) across the full corpus, escalation to stronger models only for documents the first pass flags as complex or borderline (typically single-digit percent), and checkpointed submission so re-runs never re-bill completed documents. The arithmetic at scale: a million-document matter that pencils out unaffordably at naive frontier-synchronous pricing routinely lands at a tenth of that figure under the tiered-batch structure, which moves AI-assisted review from boutique luxury to standard practice. Data paths stay inside the review platform's protected environment throughout, documented for the inevitable discovery-about-discovery motion.

Does the BAA channel itself cost us a premium versus direct APIs?

Far less than assumed, and usually nothing that matters. Azure OpenAI and Bedrock price their models comparably to direct vendor APIs (the same or within a modest spread by model and region), and both carry the cost-relevant features: prompt caching with its steep input discounts, batch endpoints at roughly half price, and a full lineup of model sizes for routing. The genuine costs of the compliant channel are slower model availability (new releases sometimes land on direct APIs first, a lag worth scheduling around rather than lamenting) and region constraints that occasionally affect pricing tiers. What we find in Philadelphia audits is that teams attribute their high bills to 'the compliance premium' when the actual driver is unoptimized engineering inside the channel: uncached context, oversized models, synchronous defaults. The audit separates those stories with numbers, because blaming the BAA for prompt bloat leaves the savings on the table and the posture unfairly maligned.

How do we justify the optimization spend to a hospital or firm budget committee?

With the same artifact those committees already trust: a baseline, a delta, and a payback period, all independently verifiable. The report structure: a month of pre-change usage instrumented and converted to the unit the committee budgets in (cost per appeal case, per reviewed document, per correspondence piece); each fix's contribution attributed in those units with the method documented; and the payback arithmetic stated plainly (engagement fee against monthly savings, typically clearing in one to three months at institutional volumes). For hospitals, the framing that lands is redirected capacity: per-case AI costs dropping 70 percent at appeal-queue volume funds the equivalent of staff positions without a hiring requisition. For firms, it is matter economics: review costs that bill through or write off either way. We also leave the instrumentation behind, so next year's committee question ('did the savings persist?') answers itself from the dashboard rather than from our slide.